Expert data and evals,
by people who do the work.
Training data, physical AI and red-teaming, from vetted experts and skilled workers.





Gemini
Claude
Codex
Cursor







Marseer AI=D6/C6AI learns from people. We make sure they’re the right ones.
- Experts vetted
- 50K+
- Specialist roles
- 30+
- Tasks delivered each day
- 80+
- QA acceptance rate
- 96%+
Sourced and vetted.
Clinicians, engineers, lawyers and skilled workers, interviewed and tested in their field before any AI work.
How we vet →Desk work, done by people who do it daily.
Engineers, clinicians, lawyers, accountants, analysts, designers and editors write, rank and check the data for the work they know.
Captured where the work happens.
Egocentric video, depth, teleoperation and annotation, recorded by people who do the task.
- 01Egocentric videoFirst-person capture of people doing real tasks.
- 02Depth and 3DStereo and LiDAR capture of real scenes.
- 03TeleoperationHuman-guided robot demonstrations, episode by episode.
- 04Video and sensor annotationSegmentation, captions, actions and keypoints.
- 05Bring your own rigYour hardware and protocol, our crews and settings.
Captured across 13 settings.
The people behind safer AI.
Clinicians, lawyers and engineers, plus housekeepers, construction crews and warehouse teams. Vetted before they touch a model.








- 01Interviewed in their field
AI voice interviews test how experts reason, not only what their CV says.
- 02Tested on real work
Field-specific work tests and a calibration set before any live task.
- 03Paid per approved task
Work that fails review is redone or rejected, and not billed to you.
- 04Re-checked against gold
Ongoing spot checks against gold answers keep quality from drifting.
Environments, models
and benchmarks.
An RL platform for agents, llm and slm model training, and benchmarking on gold datasets.
RL environments.
Sandboxes where agents and LLMs act on tasks written by vetted experts, scored by rewards those experts designed.
Read about the platformHush
Removes noise and keeps the voice, on the device.
Turn detection · edgeCue
Knows when a speaker has finished, so agents reply on time.
Working memory · PythonThredmark
Source-backed working memory for evolving AI workflows.
Document parsingParsing models
Any document, read at 99% accuracy, faster and cheaper.
Consent, provenance, control.
You know who made each item, who checked it, and where it lives.
- Consent on file
Every expert and every person on camera has signed consent before work starts.
- Provenance per item
Who made it, who reviewed it and when, attached to every delivery.
- Your region, your bucket
Delivered to your cloud storage in the region you choose.
- Confidential by default
Experts sign NDAs. Data is segregated by customer and deleted on request.
- Independent
No AI lab owns part of Saolabs. Your evaluation data is never shared.
- Certified platform
Built on Scrini AI's platform: SOC 2 Type II, ISO 27001 and ISO 27701 certified.
Start small, then scale.
A written spec, a paid pilot, then scheduled delivery. No long contract up front.
- Day 0Describe the failure
Tell us where your model breaks and what good looks like.
- Within 48 hoursSpec and quote
A written spec: experts, volume, format, QA bar and price.
- Week 1 to 2Paid pilot
A small batch with a QA report, so you judge the work first.
- OngoingScheduled delivery
Regular batches into your bucket, in your format and region.













